Vehicle Target Detection Based On UAV Image

With the development of social economy,the car utilization rate of residents in our country is increasing year by year.The frequent road congestion,traffic emergencies,traffic safety issues,etc.have affected the normal travel of residents and restricted the development of the urban economy.With the advancement of technology,intelligent transportation systems have gradually matured,and the collection of road traffic information urgently needs a stable and efficient information collection method.The vehicle target detection using aerial imagery of drones provides a stable and efficient method for the collection of road traffic information.This thesis mainly studies two vehicle target detection algorithms based on convolutional neural networks,and improves the two algorithms based on the characteristics of the data set produced by UAV aerial imagery.The main work of this paper includes the following parts.(1)A vehicle target detection data set based on drone aerial images was produced.The important premise of vehicle target detection using aerial imagery of drones using deep learning is to create vehicle targets containing a large number of aerial imagery of drones.The data divides vehicles into three categories,cars,trucks,and buses,using the open source software label Img to label the specific information of the vehicle target,the aerial image data set of the drone produced in this paper contains 2190 pictures,containing a large number of vehicle targets.(2)Realize vehicle target detection of UAV aerial imagery based on Faster R-CNN(Region-based Convolution Neural Networks).The original Faster R-CNN has a poor detection effect on small targets for vehicle detection,especially incomplete vehicle targets.Based on the characteristics of the data set made in this article,the original Faster R-CNN is improved.Based on the original Faster R-CNN,with upsampling Soft-NMS and Res Net-101 as a feature extraction network were established.The original Faster R-CNN UAV vehicle target detection and improved Faster R-CNN UAV vehicle target detection results are analyzed and compared.Compared with the original network,improved Faster R-CNN UAV aerial image vehicle target Compared with the original mAP,the detection is improved by 1.74%.(3)Realize vehicle target detection based on YOLOv3(You Look Only Once)UAV aerial imagery.This article uses the v3 version of YOLO.The original version of YOLOv3 does not have a good detection effect on small targets for vehicle detection,including incomplete vehicle targets.According to the characteristics of the data set made in this article,the original YOLOv3 is improved,replace Darknet-53 with deep-level Res Net-101 as the feature extraction network.Compared with the original network,the improved YOLOv3 UAV vehicle target detection is more adaptable to small target vehicles and incomplete target vehicles in aerial images and the detection accuracy is higher than the original mAP by 5.36%.